The next great training dataset may not be a dataset. It may be a world the model writes for itself. New paper: SPADE : Self-Play in Adaptive Synthetic Executable Environments Most agents train inside fixed worlds: curated tasks, static synthetic pools or frozen verifiers.